Democratizing Cardiovascular AI: A Multi-Collection RAG Architecture and Product Requirements for the Cardiology Intelligence Agent

Author: Adam Jones Date: March 2026 Version: 0.1.0 (Pre-Implementation) License: Apache 2.0

Part of the HCLS AI Factory -- an end-to-end precision medicine platform. https://github.com/ajones1923/hcls-ai-factory

---

Abstract

Cardiovascular disease (CVD) remains the leading cause of death globally, claiming approximately 17.9 million lives annually -- representing 32% of all deaths worldwide. Despite exponential growth in cardiovascular AI research (over 4,500 publications in 2025 alone), critical barriers persist in clinical translation: fragmented evidence across modalities, siloed data systems, lack of integrated genomic-imaging correlation, and prohibitive infrastructure costs that limit advanced cardiac AI to elite academic centers.

This paper presents the architectural design, clinical rationale, and product requirements for the Cardiology Intelligence Agent -- a multi-collection retrieval-augmented generation (RAG) system purpose-built for cardiovascular medicine. The agent will unify 12 specialized Milvus vector collections spanning cardiac imaging (echocardiography, cardiac CT, cardiac MRI, nuclear cardiology), electrophysiology (12-lead ECG, Holter monitoring, device interrogation), hemodynamics, heart failure management, valvular heart disease, preventive cardiology, interventional cardiology, and cardio-oncology -- alongside a shared genomic_evidence collection containing 3.5 million variant vectors from the HCLS AI Factory genomics pipeline.

The system extends the proven multi-collection RAG architecture established by five existing intelligence agents in the HCLS AI Factory (Precision Biomarker, Precision Oncology, CAR-T, Imaging, and Autoimmune agents), adapting it with cardiology-specific clinical workflows, risk calculators, cross-modal imaging-genomics triggers, and structured reporting aligned with ACC/AHA guidelines. Eight reference clinical workflows will cover the highest-impact cardiovascular use cases: coronary artery disease assessment, heart failure classification, valvular disease quantification, arrhythmia detection, cardiac MRI tissue characterization, stress test interpretation, preventive risk stratification, and cardio-oncology surveillance.

The agent will deploy on a single NVIDIA DGX Spark ($3,999) using BGE-small-en-v1.5 embeddings (384-dimensional, IVF_FLAT, COSINE), Claude Sonnet 4.6 for evidence synthesis, and four NVIDIA NIM microservices for on-device inference. Licensed under Apache 2.0, the platform will democratize access to integrated cardiovascular intelligence that currently requires multi-million-dollar institutional investments in informatics infrastructure.

---

Table of Contents

1. Introduction
2. The Cardiovascular Data Challenge
3. Clinical Landscape and Market Analysis
4. Existing HCLS AI Factory Architecture
5. Cardiology Intelligence Agent Architecture
6. Milvus Collection Design
7. Clinical Workflows
8. Cross-Modal Integration
9. NIM Integration Strategy
10. Knowledge Graph Design
11. Query Expansion and Retrieval Strategy
12. API and UI Design
13. Clinical Decision Support Engines
14. Reporting and Interoperability
15. Product Requirements Document
16. Data Acquisition Strategy
17. Validation and Testing Strategy
18. Regulatory Considerations
19. DGX Compute Progression
20. Implementation Roadmap
21. Risk Analysis
22. Competitive Landscape
23. Discussion
24. Conclusion
25. References

---

1. Introduction

1.1 The Cardiovascular Disease Burden

Cardiovascular disease represents the single largest cause of mortality and morbidity worldwide. According to the World Health Organization and the American Heart Association (AHA) 2025 Heart Disease and Stroke Statistics Update:

• 17.9 million deaths annually -- 32% of all global deaths
• 523 million people currently living with cardiovascular disease
• $407 billion in direct and indirect costs in the United States alone (2024)
• Ischemic heart disease and stroke are the top two causes of death globally, responsible for a combined 15.2 million deaths per year
• Heart failure affects 64.3 million people globally, with 5-year mortality rates exceeding 50%
• Atrial fibrillation prevalence has reached 59 million cases worldwide, with a projected 72 million by 2030

Despite these staggering numbers, cardiovascular AI adoption in clinical practice remains in early stages. A 2025 survey by the American College of Cardiology (ACC) found that while 89% of cardiologists believe AI will transform their practice, only 23% currently use AI-assisted tools in routine clinical decision-making. The gap between potential and adoption is driven by three factors: fragmented data, prohibitive cost, and lack of integrated systems.

1.2 The Opportunity for Integrated Cardiovascular Intelligence

The cardiovascular domain is uniquely suited for an integrated AI intelligence agent because:

1. Multi-modal data convergence: Cardiology routinely integrates imaging (echo, CT, MRI, nuclear), electrophysiology (ECG, Holter), hemodynamics (catheterization), biomarkers (troponin, BNP, lipids), genomics (familial cardiomyopathy, channelopathies), and clinical risk scores -- all of which must be synthesized for optimal patient care.

2. Established clinical guidelines: ACC/AHA guidelines provide structured frameworks for risk stratification, treatment decisions, and follow-up protocols that can be encoded into clinical decision support logic.

3. Quantitative measurement-rich: Unlike many medical specialties, cardiology generates highly quantitative data (ejection fraction percentages, valve gradients in mmHg, calcium scores in Agatston units, QTc intervals in milliseconds) that is well-suited for structured RAG retrieval and comparison.

4. Strong cross-modal triggers: A cardiac imaging finding (e.g., unexplained left ventricular hypertrophy) frequently triggers genomic workup (hypertrophic cardiomyopathy gene panel), which may lead to drug therapy optimization -- a natural fit for the HCLS AI Factory's three-stage pipeline.

5. Large addressable market: Every health system has a cardiology department. Cardiovascular AI is projected to reach $2.8 billion by 2028 (CAGR 42.3%), making this the highest-value vertical for an intelligence agent.

1.3 Our Contribution

This paper presents the complete architectural blueprint and product requirements for the Cardiology Intelligence Agent, the sixth domain-specific intelligence agent in the HCLS AI Factory platform. Our contributions include:

• A 12-collection Milvus vector schema designed for the full spectrum of cardiovascular data: imaging, electrophysiology, hemodynamics, heart failure, valvular disease, preventive cardiology, interventional procedures, cardio-oncology, guidelines, trials, devices, and literature
• Eight reference clinical workflows covering coronary artery disease, heart failure, valvular disease, arrhythmia, cardiac MRI, stress testing, preventive risk stratification, and cardio-oncology surveillance
• A cardiology knowledge graph with structured data on 30+ cardiovascular conditions, 15+ imaging modalities/protocols, 20+ cardiac biomarkers, 25+ drug classes, and 50+ ACC/AHA guideline recommendations
• Cross-modal triggers linking cardiac imaging findings to genomic workup (familial hypercholesterolemia, cardiomyopathy gene panels, channelopathies) via the shared genomic_evidence collection
• Clinical decision support engines implementing validated risk calculators (ASCVD, HEART, CHA₂DS₂-VASc, HAS-BLED, MAGGIC, EuroSCORE II)
• A comprehensive product requirements document with user stories, acceptance criteria, and implementation prioritization
• Deployment on a single NVIDIA DGX Spark ($3,999), maintaining the platform's commitment to accessible AI

---

2. The Cardiovascular Data Challenge

2.1 Data Fragmentation in Cardiovascular Medicine

Cardiovascular clinical practice generates data across at least fifteen distinct categories, each with its own structure, vocabulary, source systems, and update cadences:

1. Cardiac Imaging Literature -- PubMed abstracts, JACC imaging supplements, Circulation reviews, conference proceedings (ACC, AHA, ESC, SCMR, SCCT, ASE). Over 4,500 publications in 2025 alone on cardiovascular AI.

2. Echocardiography Data -- Chamber dimensions (LVIDd, LVIDs, LA volume), systolic function (LVEF, GLS, TAPSE, S'), diastolic function (E/A ratio, E/e', deceleration time), valve assessments (gradients, regurgitant volumes, effective orifice area), strain imaging (longitudinal, circumferential, radial). Structured per ASE guidelines.

3. Cardiac CT Data -- Coronary artery calcium scores (Agatston, volume, mass scores), CTA stenosis grading (CAD-RADS 0-5), plaque characterization (calcified, non-calcified, mixed, vulnerable), fractional flow reserve from CT (FFR-CT), myocardial perfusion CT, cardiac structural assessment.

4. Cardiac MRI Data -- Volumes and function (LVEF, RVEF, indexed volumes), tissue characterization (T1 mapping, T2 mapping, ECV), late gadolinium enhancement (LGE) patterns (ischemic vs non-ischemic), perfusion (stress, rest, MPR), feature tracking strain, 4D flow, parametric mapping.

5. Nuclear Cardiology -- SPECT MPI (stress/rest perfusion, TID ratio, transient ischemic dilation), PET MPI (MBF quantification, CFR), MUGA (LVEF for cardio-oncology), cardiac amyloid scintigraphy (Tc-99m PYP, DPD), FDG-PET for sarcoidosis/endocarditis.

6. Electrophysiology Data -- 12-lead ECG interpretation (rhythm, intervals, morphology, axis, ischemic changes), Holter/event monitor data (arrhythmia burden, HR variability), device interrogation (ICD, pacemaker, CRT), EP study results (ablation lesion sets, activation mapping), QTc monitoring.

7. Hemodynamic Data -- Right heart catheterization (PA pressures, wedge pressure, cardiac output, PVR, SVR), left heart catheterization (LVEDP, aortic valve gradient), coronary angiography (TIMI flow, stenosis percentage, FFR/iFR), structural intervention hemodynamics.

8. Heart Failure Data -- NYHA classification, ACC/AHA staging (A-D), biomarkers (NT-proBNP, BNP, hs-troponin), GDMT titration (beta-blocker, ACEi/ARB/ARNI, MRA, SGLT2i), device therapy (ICD, CRT), transplant evaluation (MELD-XI, SHFM, HFSS), LVAD management.

9. Valvular Heart Disease -- Severity grading (mild/moderate/severe per ASE criteria), hemodynamic quantification (EOA, DVI, PHT, regurgitant fraction), intervention criteria (surgical AVR vs TAVR, mitral repair vs replacement), surveillance protocols.

10. Preventive Cardiology -- Lipid panels (LDL-C, HDL-C, TG, Lp(a), ApoB), risk calculators (PCE/ASCVD, HEART, MESA, Reynolds), statin eligibility, PCSK9i criteria, coronary calcium for risk reclassification, inflammatory markers (hsCRP, IL-6), metabolic parameters.

11. Interventional Cardiology -- PCI procedural data (stent type, vessel, lesion characteristics, TIMI flow), structural intervention data (TAVR, MitraClip, WATCHMAN, PFO closure), complications, antiplatelet/anticoagulant management, CTO techniques.

12. Cardio-Oncology -- Baseline cardiac assessment protocols (pre-anthracycline, pre-immunotherapy, pre-targeted therapy), surveillance schedules (GLS monitoring, troponin trends), cardiotoxicity detection (CTRCD definitions), risk scores (HFA-ICOS), cardioprotective strategies.

13. Clinical Trials -- ClinicalTrials.gov cardiovascular entries (16,000+ active/completed), landmark trial results (PARADIGM-HF, DAPA-HF, EMPEROR-Reduced, ISCHEMIA, REVIVED-BCIS2), outcome data, subgroup analyses.

14. Cardiovascular Devices -- FDA-cleared AI/ML cardiac devices (ECG interpretation, echo measurement, CT calcium scoring, arrhythmia detection), implantable devices (ICDs, pacemakers, CRT, LVAD), wearable cardiac monitors (smartwatch ECG, continuous glucose monitors for cardiometabolic risk).

15. Cardiovascular Genomics -- Familial hypercholesterolemia (LDLR, PCSK9, APOB), hypertrophic cardiomyopathy (MYH7, MYBPC3, TNNT2, TNNI3), dilated cardiomyopathy (TTN, LMNA, RBM20), arrhythmogenic cardiomyopathy (PKP2, DSP, DSG2), channelopathies (SCN5A, KCNQ1, KCNH2, RYR2), aortopathies (FBN1, TGFBR1/2, SMAD3, ACTA2).

2.2 Why Existing Tools Fall Short

Current approaches to cardiovascular intelligence fail to address this fragmentation:

Approach	Limitation
PubMed search	Keyword-based; misses semantic connections; no cross-modal integration; no structured clinical data
UpToDate / DynaMed	Expert-curated but static; no patient-specific reasoning; no imaging or genomic data integration; subscription-based
Commercial CVIS (Solas, Lumedx)	Vendor-locked; limited to institutional data; no literature integration; $500K-$2M+ implementation
EHR-integrated CDS (Epic BestPractice)	Rule-based; cannot synthesize unstructured evidence; no imaging AI; limited to institutional data
General AI assistants	No citation provenance; hallucination risk; no structured cardiovascular data; not FDA-aligned
Imaging-only AI (Arterys, HeartFlow)	Single-modality; no genomic integration; no guideline reasoning; cloud-dependent; expensive per-scan pricing

The Cardiology Intelligence Agent addresses all six limitations simultaneously by combining multi-collection vector search, cross-modal genomic triggers, validated clinical decision support logic, and guideline-grounded LLM synthesis -- all on a $3,999 desktop device.

2.3 The Case for Multi-Collection RAG in Cardiology

A cardiologist evaluating a patient with new-onset heart failure must simultaneously consider:

• Imaging data: Echocardiographic LVEF, cardiac MRI showing late gadolinium enhancement pattern (ischemic vs non-ischemic)
• Electrophysiology: ECG showing LBBB morphology (CRT candidacy), QTc for drug safety
• Biomarkers: NT-proBNP trend, hs-troponin for myocardial injury, iron studies for IV iron candidacy
• Genomics: If age < 50 or family history, cardiomyopathy gene panel (TTN, LMNA, MYH7)
• Guidelines: ACC/AHA heart failure guidelines for GDMT initiation sequence
• Trials: DAPA-HF, EMPEROR-Reduced for SGLT2i evidence; PARADIGM-HF for ARNI
• Risk scores: MAGGIC score for prognosis, Seattle Heart Failure Model for transplant timing

No existing tool synthesizes all seven data dimensions into a single clinical narrative. A multi-collection RAG architecture -- where each data dimension has its own optimized Milvus collection with domain-specific schema fields -- enables parallel retrieval across all dimensions with a single query, followed by LLM synthesis into a coherent clinical recommendation.

---

3. Clinical Landscape and Market Analysis

3.1 Cardiovascular AI Market

The global cardiovascular AI market demonstrates exceptional growth dynamics:

Metric	Value	Source
Market size (2024)	$1.4 billion	Grand View Research
Projected size (2028)	$2.8 billion	Grand View Research
CAGR (2024-2028)	42.3%	Grand View Research
FDA-cleared cardiac AI devices (cumulative)	180+	FDA AI/ML database
Active cardiovascular AI clinical trials	450+	ClinicalTrials.gov
Annual CV AI publications	4,500+	PubMed (2025)
US cardiology practices	25,000+	ACC Census
US cardiologists	35,000+	AHA Statistics
Global cardiologists	200,000+	WHO estimates

3.2 Competitive Analysis

Competitor	Strengths	Gaps
HeartFlow (FFR-CT)	FDA-cleared, clinical validation	Single modality (CT), no genomics, cloud-only, $1,100/scan
Eko Health (ECG + auscultation)	Point-of-care, stethoscope integration	Limited to murmur/arrhythmia screening, no imaging AI
Viz.ai (Stroke + cardiac)	Real-time triage, HIPAA-compliant	Narrow scope (LVO stroke, PE), SaaS pricing
Ultromics (EchoGo)	FDA-cleared echo AI, GLS analysis	Echo-only, no cross-modal integration
Cleerly (Coronary CTA)	Plaque quantification, FDA pathway	CT-only, cloud-only, per-patient pricing
Caption Health (Echo guidance)	AI-guided acquisition, GE partnership	Acquisition guidance only, no interpretation
Tempus (Cardiology platform)	Multi-modal data, genomics integration	Proprietary, expensive, cloud-dependent

Our differentiation: The Cardiology Intelligence Agent is the only system that combines (1) multi-modal imaging AI, (2) genomic integration, (3) literature RAG, (4) guideline-aligned clinical decision support, (5) validated risk calculators, and (6) on-device deployment -- all in an open-source, $3,999 package. No competitor addresses more than two of these six dimensions.

3.3 Target Users

User Segment	Use Case	Pain Point Addressed
Community cardiologists	Evidence-based decision support	Limited access to sub-specialty expertise
Academic medical centers	Research and education	Fragmented data across systems
Heart failure programs	GDMT optimization, transplant evaluation	Manual chart review for risk stratification
Structural heart teams	TAVR/MitraClip planning	Multi-modal data synthesis
Cardio-oncology clinics	Cardiotoxicity surveillance	No integrated monitoring platform
Preventive cardiology	Risk stratification, statin eligibility	Calculator fatigue, guideline complexity
Clinical trial sites	Patient screening, endpoint adjudication	Manual eligibility assessment
Cardiovascular genomics	Variant interpretation in cardiac context	Limited cardiac-specific annotation

---

4. Existing HCLS AI Factory Architecture

4.1 Platform Overview

The HCLS AI Factory is a three-stage precision medicine platform running on NVIDIA DGX Spark:

Stage 1: Genomics Pipeline (Parabricks + DeepVariant)
    FASTQ → VCF → 3.56M annotated variants
         |
Stage 2: RAG/Chat Pipeline (Milvus + Claude)
    Variant interpretation, clinical significance
         |
Stage 3: Drug Discovery Pipeline (BioNeMo + DiffDock)
    Target → Lead compound → Docking → Drug-likeness

Five intelligence agents extend this core platform with domain-specific knowledge:

Agent	Collections	Seed Vectors	Unique Capability
Precision Biomarker	11	6,134	Biological age calculators, biomarker panels
Precision Oncology	10	609	Molecular tumor board packets, trial matching
CAR-T Intelligence	11	6,266	CAR construct comparison, manufacturing optimization
Imaging Intelligence	10	876	NIM inference, DICOM workflows, 3D segmentation
Autoimmune Intelligence	10	~500	Autoantibody panels, flare prediction

4.2 Shared Infrastructure

All agents share:

• Milvus 2.4 vector database (IVF_FLAT, COSINE, 384-dim)
• BGE-small-en-v1.5 embedding model (sentence-transformers)
• Claude Sonnet 4.6 (Anthropic) primary LLM
• genomic_evidence collection (3,561,170 variants, read-only)
• Docker Compose orchestration
• FastAPI (REST) + Streamlit (UI) pattern
• lib/hcls_common shared library (23 modules)

4.3 Proven Patterns

The Cardiology Intelligence Agent will leverage battle-tested patterns from existing agents:

Pattern	Proven In	Adaptation for Cardiology
Multi-collection parallel search	All 11 agents	12 cardiology-specific collections
Knowledge graph augmentation	CAR-T, Biomarker	Cardiac conditions, drug classes, risk factors
Query expansion maps	CAR-T (12 maps), Biomarker	Cardiology terminology (e.g., "MI" → "myocardial infarction", "STEMI", "NSTEMI", "ACS")
Comparative analysis	CAR-T, Imaging	"TAVR vs SAVR", "Amiodarone vs Sotalol"
Cross-modal genomic triggers	Imaging (Lung-RADS → EGFR)	Imaging finding → cardiomyopathy gene panel
FHIR R4 export	Imaging	DiagnosticReport with cardiac SNOMED/LOINC codes
NIM inference workflows	Imaging (4 NIMs)	Cardiac-specific NIM models
Sidebar guided tour	Imaging	Cardiology demo flow
Pre-filled example queries	Imaging, Biomarker	Cardiology-specific starter questions

---

5. Cardiology Intelligence Agent Architecture

5.1 System Diagram

+==========================================================================+
|  PRESENTATION:  Streamlit Cardiology Workbench (8536)                    |
|                 10 Tabs | Evidence | Workflows | Imaging | Risk Calcs    |
|                 FastAPI REST Server (8526)                                |
+==========================================================================+
                    |                            |
+==========================================================================+
|  INTELLIGENCE:   Cardiology RAG Engine                                   |
|                  12-collection parallel search                           |
|                  Knowledge graph (30 conditions, 20 biomarkers)          |
|                  Query expansion (15 maps, cardiology terminology)       |
|                  Comparative analysis ("TAVR vs SAVR")                   |
|                  Risk calculator engine (6 validated scores)             |
|                  GDMT optimization engine                                |
+==========================================================================+
                    |                            |
+==========================================================================+
|  INFERENCE:      NIM Services (VISTA-3D, MAISI, VILA-M3, Llama-3)       |
|                  8 Clinical Workflows:                                    |
|                  CAD | HF | Valve | Arrhythmia | CMR | Stress |         |
|                  Prevention | Cardio-Onc                                 |
+==========================================================================+
                    |                            |
+==========================================================================+
|  DATA:           Milvus 2.4 (12 cardiology collections + genomic)       |
|                  BGE-small-en-v1.5 (384-dim, IVF_FLAT, COSINE)          |
|                  PubMed, ClinicalTrials.gov, ACC/AHA guidelines          |
|                  Curated seed data (echo, CT, MRI, ECG, hemodynamics)    |
+==========================================================================+

5.2 Design Principles

1. Guideline-first reasoning: Every recommendation traces to ACC/AHA/ESC guideline evidence levels (Class I/IIa/IIb/III, LOE A/B/C)
2. Quantitative precision: Cardiac measurements retain units and reference ranges (LVEF 55-70%, E/e' < 14, QTc < 470ms)
3. Cross-modal by default: Every significant finding is checked against genomic context
4. Risk-stratified output: Severity badges and urgency routing aligned with clinical acuity
5. Graceful degradation: Full functionality in mock mode without GPU or live NIM services
6. Familiar patterns: Follows the same FastAPI + Streamlit + Milvus patterns as existing agents

5.3 Port Allocation

Port	Service
8526	FastAPI REST Server
8536	Streamlit Cardiology Workbench
19530	Milvus (shared)
8520	NIM LLM (shared)
8530	NIM VISTA-3D (shared)
8531	NIM MAISI (shared)
8532	NIM VILA-M3 (shared)

---

6. Milvus Collection Design

6.1 Index Configuration

Parameter	Value
Index type	IVF_FLAT
Metric	COSINE
nlist / nprobe	1024 / 16
Dimension	384
Embedding model	BAAI/bge-small-en-v1.5

6.2 Collection Schemas

Collection 1: cardio_literature -- ~3,000 records

Published cardiovascular research papers, reviews, and meta-analyses.

Field	Type	Description
id	VARCHAR(64)	PubMed ID or unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
title	VARCHAR(500)	Paper title
text_chunk	VARCHAR(8000)	Abstract or text section
year	INT16	Publication year
journal	VARCHAR(200)	Journal name (JACC, Circulation, EHJ, etc.)
cv_domain	VARCHAR(100)	Cardiovascular subdomain (imaging, EP, HF, valve, prevention)
modality	VARCHAR(50)	Imaging modality if applicable
study_type	VARCHAR(50)	RCT, meta-analysis, cohort, case-control, review
keywords	VARCHAR(500)	MeSH terms and author keywords

Source: PubMed E-utilities with cardiovascular MeSH filters.

Collection 2: cardio_trials -- ~500 records

Cardiovascular clinical trials from ClinicalTrials.gov and landmark trial results.

Field	Type	Description
id	VARCHAR(64)	NCT number or trial identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
title	VARCHAR(500)	Official trial title
text_summary	VARCHAR(4000)	Trial summary including results
phase	VARCHAR(20)	Phase I-IV
status	VARCHAR(30)	Active, completed, recruiting
sponsor	VARCHAR(200)	Lead sponsor
cv_domain	VARCHAR(100)	CAD, HF, valve, arrhythmia, prevention
intervention	VARCHAR(300)	Drug, device, or procedure tested
primary_endpoint	VARCHAR(300)	Primary outcome measure
enrollment	INT32	Number of participants
start_year	INT16	Year trial began
outcome_summary	VARCHAR(2000)	Key results (if completed)
landmark	BOOL	Is this a landmark trial (PARADIGM-HF, DAPA-HF, etc.)

Source: ClinicalTrials.gov V2 API with cardiovascular condition filters.

Collection 3: cardio_imaging -- ~200 records

Cardiac imaging protocols, findings, and measurements across all modalities.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Finding or protocol description
modality	VARCHAR(50)	echo, cardiac_ct, cardiac_mri, nuclear, cath
imaging_type	VARCHAR(100)	TTE, TEE, stress_echo, CTA, calcium_score, etc.
finding_category	VARCHAR(100)	Chamber size, systolic function, diastolic function, valve, plaque, LGE, perfusion
measurement_name	VARCHAR(100)	LVEF, GLS, E/e', calcium_score, stenosis_pct
measurement_value	VARCHAR(50)	Numeric value with units
reference_range	VARCHAR(100)	Normal range per guidelines
severity	VARCHAR(20)	Normal, mild, moderate, severe
guideline_source	VARCHAR(100)	ASE, SCCT, SCMR, ASNC
clinical_significance	VARCHAR(500)	Interpretation guidance

Source: Curated from ASE/SCCT/SCMR/ASNC guideline documents and reference texts.

Collection 4: cardio_electrophysiology -- ~150 records

ECG interpretation, arrhythmia classification, and electrophysiology data.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	ECG pattern or arrhythmia description
ecg_pattern	VARCHAR(100)	LBBB, RBBB, LVH, ST_elevation, AF, VT, etc.
rhythm	VARCHAR(50)	Sinus, AF, flutter, SVT, VT, VF
rate_category	VARCHAR(30)	Bradycardia, normal, tachycardia
interval	VARCHAR(50)	PR, QRS, QT, QTc
interval_value_ms	FLOAT	Numeric interval value
clinical_significance	VARCHAR(500)	What this finding means
urgency	VARCHAR(20)	Routine, urgent, emergent
differential_diagnosis	VARCHAR(500)	DDx list
management	VARCHAR(500)	Recommended next steps

Source: Curated from ACC/AHA/HRS guidelines and electrophysiology references.

Collection 5: cardio_heart_failure -- ~150 records

Heart failure classification, GDMT protocols, and management algorithms.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	HF management recommendation
hf_type	VARCHAR(30)	HFrEF, HFmrEF, HFpEF
acc_aha_stage	VARCHAR(5)	A, B, C, D
nyha_class	VARCHAR(5)	I, II, III, IV
drug_class	VARCHAR(100)	Beta-blocker, ARNI, MRA, SGLT2i, hydralazine/nitrate
drug_name	VARCHAR(100)	Specific medication
target_dose	VARCHAR(50)	Goal dose per guidelines
titration_protocol	VARCHAR(500)	How to uptitrate
evidence_level	VARCHAR(20)	Class I/IIa/IIb/III, LOE A/B/C
landmark_trial	VARCHAR(100)	Supporting trial name
contraindications	VARCHAR(500)	When not to use

Source: ACC/AHA/HFSA 2022 Heart Failure Guidelines, ESC 2023 update.

Collection 6: cardio_valvular -- ~120 records

Valvular heart disease assessment, severity grading, and intervention criteria.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Valve assessment or recommendation
valve	VARCHAR(30)	Aortic, mitral, tricuspid, pulmonic
pathology	VARCHAR(50)	Stenosis, regurgitation, prolapse, endocarditis
severity	VARCHAR(20)	Mild, moderate, severe
quantitative_criteria	VARCHAR(500)	EOA, gradient, regurgitant volume, EROA, vena contracta
intervention_criteria	VARCHAR(500)	When to intervene per guidelines
intervention_type	VARCHAR(100)	SAVR, TAVR, mitral_repair, MitraClip, TMVr
sts_risk_threshold	VARCHAR(50)	Low (<4%), intermediate (4-8%), high (>8%)
evidence_level	VARCHAR(20)	Class/LOE
surveillance_protocol	VARCHAR(500)	Follow-up echo frequency

Source: ACC/AHA 2020 Valvular Heart Disease Guidelines, ESC 2021 VHD Guidelines.

Collection 7: cardio_prevention -- ~150 records

Preventive cardiology: risk stratification, lipid management, and lifestyle intervention.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Prevention recommendation
risk_category	VARCHAR(50)	Low, borderline, intermediate, high, very_high
risk_calculator	VARCHAR(50)	PCE/ASCVD, MESA, Framingham, Reynolds
biomarker	VARCHAR(100)	LDL-C, Lp(a), ApoB, hsCRP, CAC
target_value	VARCHAR(50)	LDL < 70, LDL < 55 (very high risk)
therapy_class	VARCHAR(100)	Statin, ezetimibe, PCSK9i, bempedoic acid, inclisiran
therapy_name	VARCHAR(100)	Specific medication
evidence_level	VARCHAR(20)	Class/LOE
guideline_source	VARCHAR(100)	ACC/AHA 2018 Cholesterol, ESC 2019 Dyslipidemia
lifestyle_intervention	VARCHAR(500)	Diet, exercise, smoking, weight management

Source: ACC/AHA 2019 Prevention Guidelines, 2018 Cholesterol Guidelines.

Collection 8: cardio_interventional -- ~100 records

Interventional cardiology procedures, techniques, and outcomes data.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Procedure description or outcome data
procedure_type	VARCHAR(100)	PCI, TAVR, MitraClip, WATCHMAN, PFO_closure, CTO_PCI
indication	VARCHAR(200)	Clinical indication
technique	VARCHAR(300)	Specific approach (radial, femoral, antegrade, retrograde)
device	VARCHAR(200)	Stent, valve, clip, occluder type
success_rate	VARCHAR(50)	Procedural success rate
complication_rate	VARCHAR(50)	Major adverse event rate
antithrombotic_protocol	VARCHAR(500)	Antiplatelet/anticoagulant regimen
evidence_level	VARCHAR(20)	Class/LOE
landmark_trial	VARCHAR(100)	Supporting trial

Source: SCAI guidelines, ACC/AHA PCI and VHD guidelines.

Collection 9: cardio_oncology -- ~100 records

Cardio-oncology surveillance, cardiotoxicity detection, and cardioprotection.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Cardio-oncology recommendation
cancer_therapy	VARCHAR(100)	Anthracycline, trastuzumab, ICI, TKI, radiation
cardiotoxicity_type	VARCHAR(100)	CTRCD, myocarditis, QTc_prolongation, HTN, VTE
monitoring_protocol	VARCHAR(500)	Echo frequency, GLS thresholds, troponin schedule
gls_threshold	VARCHAR(50)	>15% relative decline = subclinical toxicity
risk_score	VARCHAR(50)	HFA-ICOS baseline risk
cardioprotective_strategy	VARCHAR(500)	Dexrazoxane, beta-blocker, ACEi, statin
when_to_hold	VARCHAR(500)	Criteria to hold/discontinue cancer therapy
evidence_level	VARCHAR(20)	Class/LOE
guideline_source	VARCHAR(100)	ESC 2022 Cardio-Oncology, ASCO

Source: ESC 2022 Cardio-Oncology Guidelines, ASCO Cardiovascular Toxicity Guidelines.

Collection 10: cardio_devices -- ~80 records

FDA-cleared cardiovascular AI devices and implantable cardiac devices.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Device description and capabilities
device_name	VARCHAR(200)	Commercial name
manufacturer	VARCHAR(200)	Company
device_category	VARCHAR(50)	AI_software, implantable, wearable
modality	VARCHAR(50)	ECG, echo, CT, MRI, wearable
clinical_task	VARCHAR(100)	Detection, measurement, prediction
regulatory_status	VARCHAR(50)	FDA_cleared, CE_marked, investigational
clearance_date	VARCHAR(20)	Approval date
performance_summary	VARCHAR(500)	Sensitivity, specificity, AUC
intended_use	VARCHAR(500)	FDA-labeled intended use statement

Source: FDA AI/ML-Enabled Medical Device database, company filings.

Collection 11: cardio_guidelines -- ~150 records

ACC/AHA/ESC/HRS clinical practice guidelines and focused updates.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Guideline recommendation
guideline_name	VARCHAR(300)	Full guideline title
organization	VARCHAR(50)	ACC/AHA, ESC, HRS, SCAI, ASE
year	INT16	Publication year
cv_domain	VARCHAR(100)	CAD, HF, valve, arrhythmia, prevention
recommendation_class	VARCHAR(10)	I, IIa, IIb, III
level_of_evidence	VARCHAR(5)	A, B-R, B-NR, C-LD, C-EO
key_recommendation	VARCHAR(2000)	Specific recommendation text
clinical_scenario	VARCHAR(500)	When this recommendation applies
related_guidelines	VARCHAR(300)	Cross-references to other guidelines

Source: ACC/AHA Practice Guidelines library, ESC Clinical Practice Guidelines.

Collection 12: cardio_hemodynamics -- ~80 records

Hemodynamic data: catheterization, pressure tracings, and derived calculations.

Field	Type	Description
id	VARCHAR(64)	Unique identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Hemodynamic scenario or reference
measurement_type	VARCHAR(100)	PCWP, PA_pressure, CO, SVR, PVR, LVEDP
normal_range	VARCHAR(50)	Expected normal values
abnormal_pattern	VARCHAR(200)	What elevation/depression indicates
associated_conditions	VARCHAR(500)	Clinical conditions causing this pattern
calculation_formula	VARCHAR(300)	How to derive (e.g., SVR = 80*(MAP-RAP)/CO)
clinical_significance	VARCHAR(500)	Interpretation guidance
catheterization_type	VARCHAR(50)	Right heart, left heart, coronary

Source: Curated from hemodynamic references and catheterization guidelines.

Collection 13: genomic_evidence -- 3,561,170 records (read-only)

Shared genomic variant collection from HCLS AI Factory Stage 1+2.

Field	Type	Description
id	VARCHAR(64)	Variant identifier
embedding	FLOAT_VECTOR(384)	BGE-small-en-v1.5 embedding
text_summary	VARCHAR(4000)	Variant annotation
gene	VARCHAR(50)	Gene symbol
clinical_significance	VARCHAR(50)	Pathogenic, likely pathogenic, VUS, benign
disease_associations	VARCHAR(1000)	Associated diseases
am_pathogenicity	FLOAT	AlphaMissense pathogenicity score

Purpose: Cross-modal triggers query this collection for cardiovascular-relevant genes (LDLR, MYH7, TTN, SCN5A, etc.).

6.3 Collection Search Weights

Collection	Weight	Rationale
Literature	0.16	Largest corpus, broadest evidence
Guidelines	0.14	Highest clinical authority
Trials	0.12	Primary evidence source
Imaging	0.10	Multi-modal imaging data
Heart Failure	0.08	Common, high-impact domain
Electrophysiology	0.08	Arrhythmia management
Valvular	0.07	Structural heart disease
Prevention	0.07	Primary/secondary prevention
Hemodynamics	0.05	Invasive assessment
Interventional	0.05	Procedural decision-making
Cardio-Oncology	0.04	Growing subspecialty
Devices	0.04	Technology landscape

6.4 Estimated Vector Counts

Collection	Seed Records	Post-Ingest Target
cardio_literature	200	3,000+ (PubMed ingest)
cardio_trials	50	500+ (ClinicalTrials.gov ingest)
cardio_imaging	200	200
cardio_electrophysiology	150	150
cardio_heart_failure	150	150
cardio_valvular	120	120
cardio_prevention	150	150
cardio_interventional	100	100
cardio_oncology	100	100
cardio_devices	80	80
cardio_guidelines	150	150
cardio_hemodynamics	80	80
Total (owned)	~1,530	~4,780+
genomic_evidence (read-only)	--	3,561,170

---

7. Clinical Workflows

7.1 Workflow Architecture

All workflows follow the established BaseImagingWorkflow pattern: preprocess → infer → postprocess → WorkflowResult. Each workflow supports full mock mode with clinically realistic synthetic results.

7.2 Eight Reference Workflows

Workflow 1: Coronary Artery Disease Assessment

Attribute	Value
Workflow ID	coronary_artery_disease
Input	Coronary CTA, calcium score CT, or catheterization data
Target Latency	< 3 minutes
Models	VISTA-3D (coronary segmentation), vessel analysis pipeline
Key Outputs	CAD-RADS classification (0-5), Agatston calcium score, per-vessel stenosis grading, plaque characterization, FFR-CT estimate
Severity Routing	CAD-RADS 4-5 → Urgent cardiology consult
Cross-Modal Trigger	CAD-RADS 4+ with age < 55 → FH gene panel (LDLR, PCSK9, APOB)
Guideline Alignment	ACC/AHA 2021 Chest Pain Guidelines, SCCT CAD-RADS 2.0

Clinical Decision Logic:

Calcium Score:
  0        → Very low risk, no further imaging needed
  1-99     → Low risk, consider statin, lifestyle
  100-399  → Moderate risk, statin indicated, stress test if symptoms
  ≥400     → High risk, coronary CTA or functional testing

CAD-RADS:
  0        → No plaque, no stenosis
  1        → 1-24% stenosis, minimal
  2        → 25-49%, mild
  3        → 50-69%, moderate → consider functional testing
  4A       → 70-99% → catheterization or FFR-CT
  4B       → Left main ≥ 50% or 3-vessel ≥ 70% → surgical evaluation
  5        → Total occlusion → viability assessment

Plaque features → High-risk if:
  - Low-attenuation plaque (< 30 HU)
  - Positive remodeling (RI > 1.1)
  - Napkin-ring sign
  - Spotty calcification

Workflow 2: Heart Failure Classification and GDMT Optimization

Attribute	Value
Workflow ID	heart_failure_gdmt
Input	Echo LVEF, BNP/NT-proBNP, clinical parameters
Target Latency	< 30 seconds
Key Outputs	HF classification (HFrEF/HFmrEF/HFpEF), ACC/AHA stage, NYHA class, GDMT recommendations with target doses, MAGGIC score
Severity Routing	Stage D or NYHA IV → Advanced HF evaluation
Cross-Modal Trigger	Age < 50 + non-ischemic DCM → cardiomyopathy gene panel (TTN, LMNA, MYH7, MYBPC3)
Guideline Alignment	ACC/AHA 2022 HF Guidelines, HFSA 2024 Update

GDMT Optimization Algorithm:

HFrEF (LVEF ≤ 40%):
  Pillar 1: Beta-blocker (carvedilol, metoprolol succinate, bisoprolol)
  Pillar 2: ARNI (sacubitril/valsartan) or ACEi/ARB
  Pillar 3: MRA (spironolactone, eplerenone)
  Pillar 4: SGLT2i (dapagliflozin, empagliflozin)

  → If African American + NYHA III-IV: Add hydralazine/isosorbide dinitrate
  → If LVEF ≤ 35% + NYHA II-III + QRS ≥ 150ms LBBB: CRT
  → If LVEF ≤ 35% + ≥ 40 days post-MI or ≥ 90 days on optimal GDMT: ICD
  → If Stage D refractory: LVAD or transplant evaluation

HFpEF (LVEF ≥ 50%):
  SGLT2i (Class I, LOE A -- EMPEROR-Preserved, DELIVER)
  Diuretics for congestion
  Treat comorbidities (HTN, AF, obesity, sleep apnea, CAD)
  GLP-1 RA if obesity + HFpEF (STEP-HFpEF)

Workflow 3: Valvular Heart Disease Quantification

Attribute	Value
Workflow ID	valvular_disease
Input	Echocardiographic measurements, clinical parameters
Target Latency	< 30 seconds
Key Outputs	Valve severity grading, intervention threshold assessment, SAVR vs TAVR recommendation for AS, STS-PROM risk score
Severity Routing	Severe symptomatic AS or MR → Heart team evaluation
Cross-Modal Trigger	Bicuspid AV + aortic dilation → aortopathy gene panel (FBN1, TGFBR1/2, SMAD3, ACTA2)
Guideline Alignment	ACC/AHA 2020 VHD Guidelines, ESC 2021 VHD Guidelines

Severity Grading Criteria (Aortic Stenosis):

Parameter	Mild	Moderate	Severe
Peak velocity (m/s)	2.0-2.9	3.0-3.9	≥ 4.0
Mean gradient (mmHg)	< 20	20-39	≥ 40
AVA (cm²)	> 1.5	1.0-1.5	< 1.0
Indexed AVA (cm²/m²)	> 0.85	0.60-0.85	< 0.60

Workflow 4: Arrhythmia Detection and Management

Attribute	Value
Workflow ID	arrhythmia_management
Input	ECG data, clinical parameters, cardiac imaging
Target Latency	< 15 seconds
Key Outputs	Rhythm classification, CHA₂DS₂-VASc score (if AF), HAS-BLED score, rate vs rhythm control recommendation, ablation candidacy
Severity Routing	VT/VF or complete heart block → Emergent
Cross-Modal Trigger	Unexplained VT in young patient → channelopathy panel (SCN5A, KCNQ1, KCNH2, RYR2)
Guideline Alignment	ACC/AHA/HRS 2023 AF Guidelines, 2017 VA/SCD Guidelines

CHA₂DS₂-VASc Calculator:

C  Congestive HF (or LVEF ≤ 40%)     +1
H  Hypertension                        +1
A₂ Age ≥ 75                            +2
D  Diabetes mellitus                   +1
S₂ Stroke/TIA/thromboembolism          +2
V  Vascular disease (PAD, prior MI)    +1
A  Age 65-74                           +1
Sc Sex category (female)               +1

Score 0 (males) or 1 (females): No anticoagulation
Score 1 (males): Consider anticoagulation
Score ≥ 2: Anticoagulation recommended (Class I)

Workflow 5: Cardiac MRI Tissue Characterization

Attribute	Value
Workflow ID	cardiac_mri_tissue
Input	Cardiac MRI data (cine, T1/T2 mapping, LGE, perfusion)
Target Latency	< 5 minutes
Models	VISTA-3D (cardiac segmentation), parametric map analysis
Key Outputs	Biventricular volumes and function, LGE pattern (ischemic vs non-ischemic), T1/T2/ECV values, perfusion defect quantification, tissue diagnosis
Severity Routing	Active myocarditis or severe biventricular failure → Urgent
Cross-Modal Trigger	Non-ischemic LGE + LV dysfunction → cardiomyopathy gene panel
Guideline Alignment	SCMR 2020 Standardized Imaging Protocols, Lake Louise Criteria

LGE Pattern Interpretation:

Pattern	Location	Diagnosis
Subendocardial / transmural	Coronary territory	Ischemic cardiomyopathy (prior MI)
Mid-wall (septal)	Interventricular septum	Dilated cardiomyopathy, myocarditis
Epicardial / patchy	Lateral wall	Myocarditis (acute/chronic)
Diffuse subendocardial	Global	Cardiac amyloidosis
RV insertion point	Anterior/posterior	Pulmonary hypertension, HCM
Asymmetric septal	Basal septum	Hypertrophic cardiomyopathy
Basal inferolateral	Mid-wall	Cardiac sarcoidosis, Chagas
No LGE + elevated T1/ECV	Global	Early-stage amyloidosis or diffuse fibrosis

Workflow 6: Stress Test Interpretation

Attribute	Value
Workflow ID	stress_test_interpretation
Input	Exercise or pharmacologic stress test data (ECG, imaging)
Target Latency	< 30 seconds
Key Outputs	Duke Treadmill Score, ischemia assessment, functional capacity (METs), hemodynamic response, stress imaging interpretation
Severity Routing	High-risk DTS or large perfusion defect → Catheterization
Guideline Alignment	ACC/AHA 2021 Chest Pain Guidelines, ASNC Stress Testing Standards

Duke Treadmill Score:

DTS = Exercise time (minutes) - (5 × max ST deviation mm) - (4 × angina index)
  Angina index: 0 = none, 1 = non-limiting, 2 = exercise-limiting

Low risk:     DTS ≥ +5    (annual mortality < 1%)
Moderate risk: DTS -10 to +4 (annual mortality 1-3%)
High risk:    DTS < -10   (annual mortality ≥ 5%)

Workflow 7: Preventive Risk Stratification

Attribute	Value
Workflow ID	preventive_risk_stratification
Input	Demographics, lipids, BP, diabetes status, smoking, family history
Target Latency	< 15 seconds
Key Outputs	10-year ASCVD risk, risk category, statin benefit group, statin intensity recommendation, risk-enhancing factors, CAC-based reclassification guidance
Cross-Modal Trigger	10-year ASCVD ≥ 7.5% + family history of premature CAD → FH gene panel (LDLR, PCSK9, APOB)
Guideline Alignment	ACC/AHA 2019 Primary Prevention, 2018 Cholesterol Guidelines

Statin Decision Framework:

1. Clinical ASCVD? → High-intensity statin (Class I)
   - If very high risk (recent ACS, multiple events): Add ezetimibe → if LDL still ≥ 55: Add PCSK9i

2. LDL ≥ 190 mg/dL? → High-intensity statin without risk calculation (Class I)
   - Consider FH genetic testing

3. Diabetes, age 40-75? → Moderate-intensity statin (Class I)
   - If multiple risk factors or 10y ASCVD ≥ 20%: High-intensity

4. 10-year ASCVD 7.5-19.9%? → Moderate-intensity statin (Class I)
   - If risk decision uncertain: Obtain CAC score
   - CAC = 0: Withhold statin (unless diabetes, FH, smoking)
   - CAC 1-99: Favors statin
   - CAC ≥ 100: Statin indicated

5. 10-year ASCVD 5-7.4%? → Consider if risk-enhancing factors present
   Risk-enhancing factors: Family hx premature ASCVD, Lp(a) ≥ 50 mg/dL,
   hsCRP ≥ 2.0, ABI < 0.9, metabolic syndrome, South Asian ancestry

Workflow 8: Cardio-Oncology Surveillance

Attribute	Value
Workflow ID	cardio_oncology_surveillance
Input	Cancer therapy regimen, baseline cardiac function, risk factors
Target Latency	< 30 seconds
Key Outputs	HFA-ICOS baseline risk category, surveillance protocol (echo frequency, biomarker schedule), GLS monitoring thresholds, cardioprotective recommendations, criteria for therapy modification
Severity Routing	LVEF decline > 10% to < 50% or GLS decline > 15% → Urgent cardio-oncology consult
Guideline Alignment	ESC 2022 Cardio-Oncology Guidelines, ASCO CV Toxicity Guidelines

Cardiotoxicity Monitoring Protocol:

Anthracycline-based chemotherapy:
  Baseline: Echo (LVEF + GLS) + troponin + NT-proBNP
  During: Echo every 2 cycles (or after cumulative dose milestones)
  After: Echo at 3, 6, 12 months post-completion, then annually
  GLS threshold: >15% relative decline from baseline = subclinical cardiotoxicity
  LVEF threshold: Decline >10% to below 50% = CTRCD

HER2-targeted therapy (trastuzumab):
  Echo every 3 months during therapy
  If LVEF 40-49%: Continue with closer monitoring + cardioprotection
  If LVEF < 40%: Hold therapy, refer cardio-oncology

Immune checkpoint inhibitors:
  Baseline ECG + troponin
  Troponin every cycle for first 4 cycles
  If troponin rise: Urgent CMR to rule out myocarditis
  ICI myocarditis mortality: 25-50% -- requires immediate therapy hold + high-dose steroids

---

8. Cross-Modal Integration

8.1 Cardiovascular Genomics Triggers

The Cardiology Intelligence Agent implements cross-modal triggers that automatically query the shared genomic_evidence collection (3.5M variants) when imaging or clinical findings suggest a genetic etiology:

Trigger Condition	Gene Panel Queried	Clinical Rationale
Non-ischemic DCM, age < 50	TTN, LMNA, RBM20, MYH7, MYBPC3, TNNT2, SCN5A, BAG3, PLN, FLNC, DSP	25-40% of DCM has genetic basis; LMNA carriers need ICD regardless of LVEF
Unexplained LVH, septal hypertrophy	MYH7, MYBPC3, TNNT2, TNNI3, TPM1, ACTC1, MYL2, MYL3, GLA (Fabry)	HCM is most common inherited cardiac disease (1:500)
Unexplained VT or cardiac arrest	SCN5A, KCNQ1, KCNH2, RYR2, CASQ2, PKP2, DSP, DSG2, DSC2	Channelopathy or ARVC screen; cascade screening for families
Premature CAD (< 55M / < 65F) or LDL > 190	LDLR, PCSK9, APOB, LDLRAP1, LIPA	FH prevalence 1:250; early detection enables prevention
Aortic root dilation > 4.5 cm	FBN1, TGFBR1, TGFBR2, SMAD3, ACTA2, MYH11, PRKG1, LOX	Heritable thoracic aortic disease; surgical thresholds differ by genotype
Cardiac amyloid (LGE + elevated T1)	TTR (transthyretin), AL amyloid genes	TTR amyloidosis is treatable with tafamidis; gene-specific therapy
Bicuspid aortic valve + coarctation	NOTCH1, GATA4/5/6, SMAD6	Congenital heart disease genetics

8.2 Imaging ↔ Genomics ↔ Drug Discovery Pipeline

Cardiac Imaging Finding (Echo, CT, MRI)
    |
    v
[Cross-Modal Trigger] -- Clinical criteria met?
    |                         |
    YES                       NO
    |                         |
    v                         v
[Query genomic_evidence]   [Standard RAG response]
(3.5M variant vectors)
    |
    v
[Variant Annotation]
ClinVar pathogenicity, AlphaMissense score
    |
    v
[HCLS AI Factory Stage 2: RAG Target ID]
Identify druggable targets from genetic findings
    |
    v
[HCLS AI Factory Stage 3: Drug Discovery]
BioNeMo molecule generation, DiffDock binding
    |
    v
[Clinical Output]
FHIR R4 DiagnosticReport with genomic enrichment

8.3 Integration with Other Agents

Integration	Direction	Data Flow
Imaging Agent	Bidirectional	Shares cardiac CT/MRI workflow results; receives DICOM routing for cardiac studies
Precision Biomarker Agent	Inbound	Receives cardiac biomarker reference ranges and age-stratified norms
Precision Oncology Agent	Inbound	Receives cancer therapy regimens for cardio-oncology surveillance
CAR-T Intelligence Agent	Inbound	Receives CRS-related cardiac toxicity data
Genomics Pipeline	Read-only	Queries genomic_evidence for cardiovascular gene variants
Drug Discovery Pipeline	Outbound	Sends confirmed genetic targets for compound screening

---

9. NIM Integration Strategy

9.1 Shared NIM Services

The Cardiology Agent will reuse the four NIM services already deployed by the Imaging Intelligence Agent:

NIM	Port	Cardiology Application
VISTA-3D	8530	Cardiac chamber segmentation, coronary artery segmentation, pericardial fat quantification
MAISI	8531	Synthetic cardiac CT generation for training, rare pathology simulation
VILA-M3	8532	Cardiac image interpretation, echo measurement validation, ECG pattern recognition
Llama-3 8B	8520	Clinical reasoning fallback when Claude API unavailable

9.2 VISTA-3D Cardiac Applications

VISTA-3D's 132 anatomical classes include cardiac structures suitable for:

• Chamber volumetrics: LV, RV, LA, RA segmentation for volume and EF calculation
• Myocardial segmentation: LV myocardium for mass calculation and wall motion assessment
• Pericardial analysis: Pericardial effusion detection and quantification
• Great vessel assessment: Aortic root and ascending aorta measurement
• Coronary segmentation: Combined with vessel centerline analysis for stenosis grading

9.3 Fallback Logic

Request → NIM available? → Yes → Real NIM inference
                        → No  → Cloud NIM? → Yes → NVIDIA Cloud API
                                            → No  → Mock enabled? → Yes → Synthetic result
                                                                   → No  → Error

---

10. Knowledge Graph Design

10.1 Graph Structure

The cardiology knowledge graph will contain structured entries across six dictionaries:

Dictionary	Entries	Content
Cardiovascular Conditions	~35	ICD-10 codes, diagnostic criteria, severity classification, imaging characteristics, guideline-based management
Cardiac Biomarkers	~25	Reference ranges, clinical cutoffs, kinetics, diagnostic/prognostic value
Drug Classes	~30	Mechanism, target dose, titration, contraindications, landmark trials
Risk Calculators	~8	Input variables, formula, risk categories, guideline recommendations per category
Imaging Protocols	~20	Modality, protocol parameters, normal values, abnormal patterns
Cardiovascular Genes	~50	Gene symbol, chromosome, associated conditions, inheritance, prevalence, clinical actionability

10.2 Example Knowledge Graph Entries

Condition: Hypertrophic Cardiomyopathy (HCM)

{
    "name": "Hypertrophic Cardiomyopathy",
    "icd10": "I42.1",
    "aliases": ["HCM", "HOCM", "IHSS", "hypertrophic obstructive cardiomyopathy"],
    "prevalence": "1:500 (most common inherited cardiac disease)",
    "inheritance": "Autosomal dominant, variable penetrance",
    "diagnostic_criteria": {
        "wall_thickness": "≥ 15 mm in any segment (or ≥ 13 mm with family history)",
        "lvot_gradient": "≥ 30 mmHg at rest or provocation = obstructive",
        "mri_lge": "Patchy mid-wall, RV insertion points, extensive LGE = worse prognosis"
    },
    "risk_stratification": {
        "model": "ACC/AHA 2024 HCM Guidelines SCD Risk Calculator",
        "high_risk_features": [
            "Family history of SCD from HCM",
            "Unexplained syncope",
            "Massive LVH (≥ 30 mm)",
            "NSVT on Holter",
            "Abnormal BP response to exercise",
            "Extensive LGE (≥ 15% LV mass)",
            "LVEF < 50%",
            "Apical aneurysm"
        ]
    },
    "genes": ["MYH7", "MYBPC3", "TNNT2", "TNNI3", "TPM1", "ACTC1", "MYL2", "MYL3"],
    "treatment": {
        "medical": "Beta-blockers (first-line), verapamil, disopyramide, mavacamten",
        "interventional": "Septal myectomy (surgical) or alcohol septal ablation",
        "device": "ICD if high SCD risk"
    },
    "imaging_modalities": ["echo", "cardiac_mri", "cardiac_ct"],
    "cross_modal_trigger": True
}

Biomarker: NT-proBNP

{
    "name": "NT-proBNP",
    "full_name": "N-terminal pro-B-type natriuretic peptide",
    "loinc_code": "33762-6",
    "reference_ranges": {
        "age_lt_50": {"normal": "< 300 pg/mL", "grey_zone": "300-450", "elevated": "> 450"},
        "age_50_75": {"normal": "< 300 pg/mL", "grey_zone": "300-900", "elevated": "> 900"},
        "age_gt_75": {"normal": "< 300 pg/mL", "grey_zone": "300-1800", "elevated": "> 1800"}
    },
    "clinical_use": "Heart failure diagnosis and prognosis, GDMT titration monitoring",
    "kinetics": "Half-life ~120 min, less affected by obesity than BNP, not cleared by neprilysin",
    "confounders": ["Renal dysfunction (↑)", "Obesity (↓)", "AF (↑)", "Pulmonary disease (↑)"],
    "guideline_thresholds": {
        "rule_out_hf": "< 300 pg/mL (NPV > 98%)",
        "hospitalized_hf": "Admission > 1000 associated with higher mortality",
        "gdmt_success": "> 30% reduction on optimal therapy"
    }
}

---

11. Query Expansion and Retrieval Strategy

11.1 Cardiology-Specific Query Expansion Maps

Fifteen domain-specific expansion maps will map cardiovascular terminology to related terms:

Map	Keywords → Terms	Example
Coronary Disease	25 → 180	"heart attack" → myocardial infarction, MI, STEMI, NSTEMI, ACS, acute coronary syndrome, troponin elevation
Heart Failure	20 → 150	"weak heart" → heart failure, HFrEF, HFpEF, reduced ejection fraction, systolic dysfunction, LVEF
Arrhythmia	20 → 140	"irregular heartbeat" → atrial fibrillation, AF, AFib, arrhythmia, flutter, SVT, palpitations
Valvular	15 → 100	"heart valve" → aortic stenosis, AS, mitral regurgitation, MR, TAVR, valve replacement
Lipids	15 → 100	"cholesterol" → LDL-C, HDL-C, triglycerides, statin, PCSK9, hyperlipidemia, dyslipidemia
Imaging Echo	15 → 90	"heart ultrasound" → echocardiogram, echo, TTE, TEE, LVEF, GLS, diastolic function
Imaging CT	10 → 70	"heart scan" → cardiac CT, CTA, calcium score, CAD-RADS, coronary CTA
Imaging MRI	10 → 70	"cardiac MRI" → CMR, LGE, T1 mapping, ECV, tissue characterization, parametric mapping
Electrophysiology	15 → 100	"ECG" → electrocardiogram, 12-lead, rhythm strip, QTc, ST segment, LBBB
Heart Failure Drugs	15 → 90	"entresto" → sacubitril/valsartan, ARNI, neprilysin inhibitor, PARADIGM-HF
Structural	10 → 60	"TAVR" → transcatheter aortic valve replacement, TAVI, Edwards SAPIEN, Medtronic CoreValve
Prevention	10 → 70	"heart risk" → ASCVD, cardiovascular risk, PCE, risk calculator, CAC score
Cardio-Oncology	10 → 60	"chemo heart" → cardiotoxicity, CTRCD, anthracycline, trastuzumab, GLS monitoring
Hemodynamics	10 → 60	"heart pressures" → right heart cath, Swan-Ganz, PCWP, wedge pressure, PVR, cardiac output
Devices	10 → 60	"pacemaker" → ICD, CRT, CRT-D, cardiac resynchronization, LVAD, defibrillator

Total: 200 keywords → ~1,400 expanded terms

11.2 Comparative Analysis Detection

The RAG engine auto-detects comparative queries using the same pattern established in the CAR-T and Imaging agents:

Trigger patterns: "vs", "versus", "compared to", "compare", "difference between", "better than"

Cardiology-specific comparisons:

Comparison	Clinical Relevance
"TAVR vs SAVR"	Aortic stenosis intervention strategy by risk category
"Amiodarone vs Sotalol"	AF rate/rhythm control
"DOAC vs Warfarin"	Anticoagulation strategy for AF
"PCI vs CABG"	Revascularization strategy for multivessel CAD
"CT vs MRI for cardiomyopathy"	Imaging modality selection
"Metoprolol vs Carvedilol"	Beta-blocker selection in HF
"Entresto vs Enalapril"	RAAS inhibition strategy in HFrEF
"Ablation vs Drugs for AF"	Rhythm control strategy

---

12. API and UI Design

12.1 FastAPI Endpoints (Port 8526)

Method	Path	Purpose
GET	/health	Service health, collection stats, NIM status
GET	/collections	List all collections with vector counts
POST	/query	Full RAG query with evidence synthesis
POST	/search	Evidence-only search (no LLM)
POST	/api/ask	Chat-style question answering
POST	/find-related	Cross-collection entity linking
GET	/workflows	List available clinical workflows
POST	/workflow/{name}/run	Execute a clinical workflow
GET	/demo-cases	List pre-loaded demo cases
POST	/demo-cases/{id}/run	Run a demo case
POST	/risk/calculate	Calculate validated risk score
POST	/gdmt/optimize	GDMT optimization recommendation
POST	/protocol/recommend	Imaging protocol recommendation
POST	/reports/generate	Generate report (markdown, JSON, PDF, FHIR)
GET	/knowledge/stats	Knowledge graph statistics
GET	/metrics	Prometheus-compatible metrics
GET	/nim/status	NIM service availability

12.2 Streamlit UI (Port 8536)

Ten-tab interface following the Imaging Agent's 9-tab pattern:

Tab	Purpose
Evidence Explorer	Multi-collection RAG Q&A with evidence citations, pre-filled cardiology example queries, Plotly donut chart for collection contribution
Workflow Runner	11 clinical workflows with pre-loaded demo cases, annotated cardiac images, 6-step pipeline animation, download buttons
Cardiac Imaging	Echo, CT, MRI, nuclear imaging gallery with AI annotations, before/after toggle, 3D volume viewer
Risk Calculators	Interactive ASCVD, HEART, CHA₂DS₂-VASc, HAS-BLED, MAGGIC, EuroSCORE II calculators with guideline recommendations
GDMT Optimizer	Heart failure medication optimizer: input LVEF, labs, vitals → get pillar-by-pillar titration plan
Device & AI Ecosystem	80+ FDA-cleared cardiac AI devices, searchable by modality and clinical task
Protocol Advisor	Patient-specific cardiac imaging protocol recommendations
Reports & Export	Markdown, JSON, NVIDIA-branded PDF, FHIR R4 DiagnosticReport export
Patient 360	Cross-modal cardiac dashboard: imaging + genomics + biomarkers + risk scores with interactive Plotly network graph
Guidelines & Trials	ACC/AHA guideline browser, landmark trial summaries, evidence level filtering

12.3 Sidebar

Section	Controls
Guided Tour	10-step demo flow with dismiss button
NIM Services	2x2 status grid (VISTA-3D, MAISI, VILA-M3, LLM)
Collection Stats	Metric widgets for each of 12 collections
Filters	CV domain dropdown, modality dropdown, year range slider
Collections to Search	Individual collection toggle checkboxes
Demo Mode	Load demo patient button

12.4 Demo Cases

ID	Title	Workflow	Key Features
DEMO-001	Acute Chest Pain: STEMI Evaluation	coronary_artery_disease	ECG ST-elevation, troponin kinetics, cath lab activation
DEMO-002	New Heart Failure: GDMT Initiation	heart_failure_gdmt	LVEF 25%, NYHA III, 4-pillar GDMT plan
DEMO-003	Severe Aortic Stenosis: TAVR Evaluation	valvular_disease	AVA 0.8 cm², mean gradient 48 mmHg, STS risk
DEMO-004	New-Onset AF: Stroke Prevention	arrhythmia_management	CHA₂DS₂-VASc 4, DOAC recommendation, rate control
DEMO-005	Anthracycline Cardiotoxicity Surveillance	cardio_oncology_surveillance	Pre-chemo baseline, GLS monitoring protocol

---

13. Clinical Decision Support Engines

13.1 Validated Risk Calculators

The agent will implement six validated cardiovascular risk calculators:

Calculator	Use Case	Inputs	Output
ASCVD (PCE)	10-year atherosclerotic CVD risk	Age, sex, race, TC, HDL, SBP, DM, smoking, HTN Rx	Risk %, statin recommendation
HEART Score	Chest pain risk stratification	History, ECG, age, risk factors, troponin	Score 0-10, risk category
CHA₂DS₂-VASc	AF stroke risk	CHF, HTN, age, DM, stroke, vascular, sex	Score 0-9, anticoagulation recommendation
HAS-BLED	AF bleeding risk on anticoagulation	HTN, renal/liver, stroke, bleeding, labile INR, elderly, drugs/alcohol	Score 0-9, risk category
MAGGIC	HF prognosis	Age, sex, LVEF, NYHA, SBP, BMI, creatinine, DM, COPD, HF duration, smoking, meds	1- and 3-year mortality
EuroSCORE II	Cardiac surgery risk	Age, sex, renal function, extracardiac arteriopathy, poor mobility, prior cardiac surgery, COPD, active endocarditis, critical preop state, DM on insulin, NYHA, CCS angina, LVEF, recent MI, PA pressure, urgency, weight of procedure, thoracic aorta surgery	Operative mortality %

13.2 GDMT Optimization Engine

The GDMT optimizer implements the ACC/AHA 2022 Heart Failure Guidelines algorithm:

class GDMTOptimizer:
    """Heart failure guideline-directed medical therapy optimizer.

    Implements the 4-pillar HFrEF GDMT framework with:
    - Current medication assessment
    - Target dose calculation
    - Titration schedule recommendation
    - Contraindication checking
    - Drug interaction screening
    - Lab monitoring requirements (K+, Cr, eGFR)
    """

    def optimize(
        self,
        lvef: float,
        nyha_class: int,
        current_meds: List[Medication],
        labs: LabValues,
        vitals: Vitals,
        comorbidities: List[str],
    ) -> GDMTRecommendation:
        ...

---

14. Reporting and Interoperability

14.1 Export Formats

Format	Use Case	Standards
Markdown	Clinical narrative, consultation notes	--
JSON	Programmatic consumption, dashboards	--
PDF	NVIDIA-themed clinical documentation	ReportLab
FHIR R4	EHR integration, interoperability	SNOMED CT, LOINC, DICOM

14.2 FHIR R4 Cardiovascular Coding

Element	Code System	Example Codes
Findings	SNOMED CT	22298006 (MI), 84114007 (HF), 49436004 (AF), 60234000 (AS)
Observations	LOINC	10230-1 (LVEF), 33762-6 (NT-proBNP), 2093-3 (Total cholesterol), 18262-6 (LDL)
Procedures	SNOMED CT	232717009 (CABG), 26212005 (PCI), 448076004 (TAVR)
Medications	RxNorm	Sacubitril/valsartan, carvedilol, dapagliflozin, apixaban
Imaging Studies	DICOM	US (echo), CT, MR, NM (nuclear)
Risk Scores	Custom extension	ASCVD %, CHA₂DS₂-VASc, HEART, MAGGIC

14.3 Structured Report Sections

Cardiovascular reports will follow ACC/AHA reporting standards:

1. Clinical Context -- Indication, relevant history, medications
2. Findings -- Organized by organ system (LV, RV, valves, great vessels, coronary)
3. Measurements -- Quantitative data with reference ranges and severity grading
4. Impression -- Synthesized clinical interpretation
5. Recommendations -- Guideline-based next steps with evidence levels
6. Genomic Enrichment -- Cross-modal genetic findings (if triggered)
7. Risk Scores -- Calculated risk stratification with category
8. Citations -- Evidence sources with relevance scores

---

15. Product Requirements Document

15.1 Product Vision

Vision Statement: Enable any cardiologist, anywhere, to access integrated cardiovascular intelligence combining imaging AI, genomic analysis, guideline-based decision support, and evidence synthesis -- on a single $3,999 device.

Target Users: Community cardiologists, academic cardiology fellows, heart failure programs, structural heart teams, cardio-oncology clinics, preventive cardiology practices, clinical trial sites.

15.2 User Stories

Epic 1: Evidence-Based Clinical Queries

ID	User Story	Priority	Acceptance Criteria
US-001	As a cardiologist, I want to ask clinical questions and receive evidence-grounded answers with citations, so that I can make informed decisions.	P0	Query returns answer with ≥3 citations from ≥2 collections; response time < 30 sec
US-002	As a cardiologist, I want comparative analysis ("TAVR vs SAVR") with structured tables, so that I can compare treatment options.	P0	Comparative query auto-detected; side-by-side evidence display; structured comparison table
US-003	As a fellow, I want pre-filled example queries for common scenarios, so that I can learn the system quickly.	P1	≥4 clickable example queries; each returns relevant results
US-004	As a researcher, I want to filter by CV domain, modality, and year, so that I can narrow my evidence search.	P1	Sidebar filters applied to all queries; results reflect applied filters

Epic 2: Clinical Workflows

ID	User Story	Priority	Acceptance Criteria
US-005	As a cardiologist, I want to run CAD assessment with calcium score and CTA data, so that I get CAD-RADS classification and management recommendations.	P0	Workflow returns CAD-RADS 0-5, per-vessel stenosis, plaque characterization, guideline recommendation
US-006	As an HF specialist, I want GDMT optimization for a patient with HFrEF, so that I get a 4-pillar titration plan.	P0	All 4 GDMT pillars assessed; current vs target dose displayed; titration schedule provided
US-007	As a structural heart team member, I want valve severity grading from echo measurements, so that I can assess intervention criteria.	P0	Severity grading matches ASE criteria; SAVR vs TAVR recommendation based on STS risk
US-008	As an EP physician, I want arrhythmia classification with CHA₂DS₂-VASc, so that I can manage AF stroke risk.	P0	CHA₂DS₂-VASc calculated correctly; anticoagulation recommendation per guidelines
US-009	As a cardio-oncologist, I want surveillance protocol generation, so that I know when to order echo and biomarkers.	P1	Protocol specific to cancer therapy; GLS and LVEF thresholds defined; cardioprotection recommendations

Epic 3: Risk Calculators

ID	User Story	Priority	Acceptance Criteria
US-010	As a preventive cardiologist, I want to calculate 10-year ASCVD risk, so that I can determine statin eligibility.	P0	Correct PCE calculation; risk category assignment; statin intensity recommendation
US-011	As an ED physician, I want HEART score for chest pain, so that I can risk-stratify patients.	P0	Score 0-10 calculated; risk category (low/mod/high); disposition recommendation
US-012	As a cardiologist, I want interactive risk calculator forms, so that I can enter patient data directly in the UI.	P1	Form inputs with validation; real-time score update; guideline recommendation display

Epic 4: Cross-Modal Integration

ID	User Story	Priority	Acceptance Criteria
US-013	As a cardiomyopathy specialist, I want genetic triggers from cardiac imaging, so that DCM patients automatically get gene panel recommendations.	P0	Non-ischemic DCM + age < 50 triggers gene panel; genomic hits displayed in Patient 360
US-014	As a cardiologist, I want a Patient 360 view combining imaging, genomics, biomarkers, and risk scores, so that I see the complete picture.	P1	Interactive Plotly network graph; cross-modal connections displayed; drill-down to evidence

Epic 5: Reporting and Export

ID	User Story	Priority	Acceptance Criteria
US-015	As a cardiologist, I want to export clinical reports as PDF, so that I can include them in patient records.	P0	NVIDIA-branded PDF with clinical question, analysis, evidence, risk scores; download button
US-016	As a health IT engineer, I want FHIR R4 DiagnosticReport export, so that findings integrate with our EHR.	P1	Valid FHIR R4 Bundle; SNOMED CT + LOINC coded; passes FHIR validator
US-017	As a cardiologist, I want 4 export formats (Markdown, JSON, PDF, FHIR), so that I can choose the right format for my use case.	P1	All 4 formats functional; consistent content across formats

Epic 6: Demo and Presentation

ID	User Story	Priority	Acceptance Criteria
US-018	As a demo presenter, I want 5 pre-loaded demo cases, so that I can run the demo without entering data.	P0	5 demo cases selectable from dropdown; each runs in < 30 seconds; realistic clinical output
US-019	As a new user, I want a sidebar guided tour, so that I understand the 10-tab interface quickly.	P1	Expandable tour with numbered steps; dismiss button; persists across session
US-020	As a demo presenter, I want a cardiac imaging gallery, so that I can show impressive AI-annotated images.	P1	Echo, CT, MRI images with AI overlays; before/after toggle; 3D slice viewer

15.3 Non-Functional Requirements

Requirement	Target	Rationale
RAG query latency	< 30 seconds end-to-end	Acceptable for clinical consultation workflow
Risk calculator latency	< 5 seconds	Near-instantaneous for interactive use
Workflow execution (mock)	< 10 seconds	Responsive demo experience
Availability	99.5% uptime	Clinical support tool (not life-critical)
Memory footprint	< 16 GB (agent only)	Coexist with other agents on 128 GB DGX Spark
Seed data completeness	1,500+ records across 12 collections	Sufficient for meaningful RAG retrieval
Unit test coverage	> 80%	Reliable development cycle
FHIR R4 compliance	Passes HL7 FHIR Validator	Interoperability requirement

15.4 Prioritization Matrix

Phase	Features	Timeline
Phase 1 (MVP)	RAG engine (12 collections), Evidence Explorer, 3 workflows (CAD, HF, arrhythmia), 3 risk calculators (ASCVD, CHA₂DS₂-VASc, HEART), 3 demo cases, PDF export	4-6 weeks
Phase 2 (Complete)	All 11 workflows, all 6 calculators, GDMT optimizer, FHIR R4 export, cardiac imaging gallery, Patient 360, all 5 demo cases	4-6 weeks
Phase 3 (Polish)	Guided tour, pipeline animation, cross-modal triggers, network graph, Guidelines & Trials tab, benchmark validation	2-3 weeks

---

16. Data Acquisition Strategy

16.1 Automated Ingest Pipelines

Source	Collection(s)	Method	Update Cadence
PubMed (NCBI E-utilities)	cardio_literature	MeSH-filtered abstract retrieval	Weekly
ClinicalTrials.gov (V2 API)	cardio_trials	Cardiovascular condition filter	Weekly
ACC/AHA Guidelines PDFs	cardio_guidelines	Manual curation + embedding	Per guideline update
ASE/SCCT/SCMR references	cardio_imaging	Manual curation	Per guideline update

16.2 Curated Seed Data

Collection	Records	Curation Source
cardio_imaging	200	ASE, SCCT, SCMR, ASNC guidelines and reference texts
cardio_electrophysiology	150	ACC/AHA/HRS guidelines, EP textbooks
cardio_heart_failure	150	ACC/AHA 2022 HF Guidelines, landmark trial data
cardio_valvular	120	ACC/AHA 2020 VHD Guidelines
cardio_prevention	150	ACC/AHA 2019 Prevention, 2018 Cholesterol Guidelines
cardio_interventional	100	SCAI guidelines, procedural references
cardio_oncology	100	ESC 2022 Cardio-Oncology Guidelines
cardio_devices	80	FDA AI/ML database, manufacturer data
cardio_hemodynamics	80	Catheterization references and guidelines

16.3 PubMed Search Strategy

MeSH Terms:
  "Cardiovascular Diseases"[MeSH] OR "Heart Diseases"[MeSH] OR
  "Coronary Artery Disease"[MeSH] OR "Heart Failure"[MeSH] OR
  "Atrial Fibrillation"[MeSH] OR "Cardiomyopathies"[MeSH]

AND ("Artificial Intelligence"[MeSH] OR "Machine Learning"[MeSH] OR
     "Deep Learning"[MeSH] OR "Neural Networks"[MeSH])

Filters: Published 2018-2026, English, Humans
Expected: 3,000-5,000 abstracts

---

17. Validation and Testing Strategy

17.1 Unit Tests

Test Category	Target Count	Coverage
Collection schemas	36	12 collections × 3 tests each
Risk calculators	48	6 calculators × 8 test cases each
GDMT optimizer	30	Multiple HF phenotypes and medication combos
Workflow logic	40	11 workflows × 5 test cases each
RAG engine	20	Query expansion, scoring, synthesis
Knowledge graph	15	Entity lookup, alias resolution
API endpoints	30	All REST endpoints
FHIR R4 export	15	Schema validation, coding accuracy
Cross-modal triggers	12	All trigger conditions
Total	~250

17.2 Clinical Validation

Validation Type	Method	Target
Risk calculator accuracy	Compare against published validation cohorts	< 1% deviation from reference implementations
GDMT recommendations	Review by board-certified HF cardiologist	95%+ guideline concordance
Severity grading	Compare against ASE/ACC criteria	100% match for standard inputs
Guideline citation accuracy	Verify against source guidelines	100% accurate citations
FHIR R4 compliance	HL7 FHIR Validator	Zero validation errors

17.3 End-to-End Validation Checks

Check	Criteria
Health endpoint	Returns status=healthy, all 12 collections with non-zero counts
RAG query	Returns answer with ≥3 citations in < 30 seconds
Risk calculator	All 6 calculators return correct results for known inputs
Workflow execution	All 11 workflows complete in mock mode
FHIR export	Valid R4 Bundle passes FHIR validator
PDF export	Generates downloadable PDF with NVIDIA branding
Demo cases	All 5 demo cases execute successfully
Cross-modal trigger	Genomic query fires for qualifying conditions
Comparative analysis	"TAVR vs SAVR" produces structured comparison

---

18. Regulatory Considerations

18.1 Intended Use Classification

The Cardiology Intelligence Agent is a clinical decision support (CDS) tool intended for use by licensed healthcare professionals. It is designed to:

• Retrieve and synthesize published cardiovascular evidence
• Calculate validated clinical risk scores
• Generate guideline-based recommendations
• Assist with (not replace) clinical decision-making

18.2 FDA CDS Exemption Criteria (21st Century Cures Act)

Under the 21st Century Cures Act, software functions meeting all four criteria are exempt from FDA device regulation:

Criterion	Assessment
Not intended to acquire, process, or analyze a medical image or signal	Met -- RAG and risk calculation; does not process raw medical images or signals
Intended for displaying, analyzing, or printing medical information	Met -- Displays evidence and calculations
Intended for use by healthcare professional	Met -- Designed for cardiologists
Healthcare professional does not primarily rely on the software	Met -- Provides recommendations for review, not autonomous decisions

Assessment: The core RAG and risk calculator functions likely qualify for CDS exemption. Clinical workflows involving NIM-based image analysis would require separate regulatory consideration.

18.3 Disclaimers

All outputs will include standard disclaimers:

This tool is for clinical decision support only and does not replace professional medical judgment. All findings require verification by a qualified cardiologist. Risk calculations are estimates based on population-level data and may not apply to individual patients. Not FDA-cleared for autonomous clinical decision-making. For research and educational purposes only.

---

19. DGX Compute Progression

Phase	Hardware	Price	Scope
Phase 1 -- Proof Build	DGX Spark	$3,999	All 11 workflows (mock/cloud NIM), 12 collections, 6 risk calculators
Phase 2 -- Departmental	1-2x DGX B200	$500K-$1M	Full NIM stack, live echo/CT/MRI processing, PACS integration
Phase 3 -- Multi-Site	4-8x DGX B200	$2M-$4M	NVIDIA FLARE federated learning across sites, population analytics
Phase 4 -- AI Factory	DGX SuperPOD	$7M-$60M+	Thousands concurrent studies, real-time ICU monitoring, national registries

---

20. Implementation Roadmap

Phase 1: Foundation (Weeks 1-6)

Week	Deliverable
1-2	Repository scaffolding, Pydantic models, settings, Docker Compose, 12 collection schemas
3-4	Seed data curation (1,530 records), ingest pipelines (PubMed, ClinicalTrials.gov), embedding generation
5-6	RAG engine (parallel search, weighted scoring, query expansion, Claude synthesis), Evidence Explorer tab

Phase 2: Clinical Intelligence (Weeks 7-12)

Week	Deliverable
7-8	3 priority workflows (CAD, HF, arrhythmia), risk calculators (ASCVD, CHA₂DS₂-VASc, HEART), demo cases
9-10	Remaining 5 workflows (valve, CMR, stress, prevention, cardio-onc), remaining calculators (HAS-BLED, MAGGIC, EuroSCORE II)
11-12	GDMT optimizer, cross-modal triggers, FHIR R4 export, PDF reports

Phase 3: UI and Polish (Weeks 13-16)

Week	Deliverable
13-14	10-tab Streamlit UI, cardiac imaging gallery, Patient 360 network graph, Guidelines & Trials browser
15-16	Sidebar guided tour, pipeline animation, pre-filled examples, 5 demo cases finalized, documentation

Phase 4: Integration and Validation (Weeks 17-18)

Week	Deliverable
17	Integration with Imaging Agent (shared cardiac workflows), cross-agent triggers, Docker Compose integration
18	Clinical validation, end-to-end testing (250+ unit tests), documentation publication (demo guide, design doc, project bible)

---

21. Risk Analysis

Risk	Probability	Impact	Mitigation
Risk calculator implementation errors	Medium	High	Validate against published reference implementations; extensive unit testing
GDMT logic complexity	Medium	Medium	Start with HFrEF only (best-defined); expand to HFmrEF/HFpEF later
Insufficient seed data for niche collections	Medium	Medium	Focus on high-impact collections first; iteratively expand
NIM cardiac segmentation accuracy	Low	Medium	Mock mode fallback; VISTA-3D already supports cardiac structures
Guideline updates during development	Low	Low	Modular guideline collection; easy to update individual recommendations
Memory pressure with 12 collections + existing agents	Low	Medium	BGE-small (384-dim) is compact; Milvus handles multi-collection efficiently
Claude API rate limits during demo	Low	High	Llama-3 NIM fallback; response caching for demo scenarios

---

22. Competitive Landscape

22.1 Positioning

The Cardiology Intelligence Agent occupies a unique position in the cardiovascular AI landscape:

                    Multi-Modal Integration
                           ↑
                           |
                    [Cardiology Agent]
                           |
         Cloud-only ←------+------→ On-device
                           |
                    [HeartFlow, Cleerly]
                           |
                           ↓
                    Single-Modality

No existing product combines:

1. Multi-modal cardiac imaging AI (echo + CT + MRI + nuclear + ECG)
2. Genomic integration with cross-modal triggers
3. RAG-grounded evidence synthesis with citations
4. Validated risk calculators (ASCVD, HEART, CHA₂DS₂-VASc)
5. GDMT optimization engine
6. On-device deployment ($3,999)
7. Open-source (Apache 2.0)

22.2 Defensibility

Advantage	Defensibility
Multi-collection RAG architecture	High -- 11 agents prove the pattern across all major clinical domains
Cross-modal genomic triggers	High -- unique to HCLS AI Factory; requires integrated platform
On-device inference	High -- DGX Spark + NIM stack is NVIDIA-exclusive
Open source	Medium -- community contribution, institutional customization
Guideline-aligned CDS	Low -- guidelines are public; competitors can implement
Clinical validation	Medium -- requires domain expertise and clinical partnerships

---

23. Discussion

23.1 Why Cardiology Is the Right Next Agent

Of all medical specialties, cardiology presents the strongest case for an integrated intelligence agent:

1. Data richness: No other specialty routinely integrates as many data modalities (imaging, EP, hemodynamics, biomarkers, genomics, risk scores) into a single clinical encounter.

2. Quantitative precision: Cardiovascular medicine is measurement-driven. LVEF, calcium scores, valve gradients, QTc intervals, and risk percentages are all inherently structured data -- ideal for RAG retrieval and clinical decision support.

3. Guideline density: The ACC/AHA and ESC produce more clinical practice guidelines than any other medical specialty. These guidelines are highly structured (Class I-III, LOE A-C) and directly mappable to clinical decision support rules.

4. Clinical workflow fit: Cardiologists already use structured reporting, risk calculators, and decision algorithms. The Cardiology Intelligence Agent enhances rather than disrupts existing workflow.

5. Market size: Cardiovascular AI is the largest segment of the medical AI market ($2.8B projected by 2028). Every health system has cardiologists. The addressable market is essentially universal.

6. Existing foundation: The HCLS AI Factory already has cardiac CT workflows (DEMO-003 in the Imaging Agent), coronary segmentation via VISTA-3D, and cardiovascular genomics in the shared variant collection. The Cardiology Agent builds on existing infrastructure rather than starting from scratch.

23.2 Limitations

1. Mock mode inference: Phase 1 uses clinically realistic mock results rather than live model inference. This is appropriate for proof-of-concept and demo but requires GPU deployment for clinical validation.

2. Guideline currency: Clinical guidelines are updated periodically. The agent requires manual curation to incorporate focused updates, although the modular collection architecture makes updates straightforward.

3. Risk calculator applicability: Validated risk calculators (PCE, HEART) were developed in specific populations and may not generalize equally across all demographic groups. The agent should display appropriate caveats.

4. Not a replacement for clinical judgment: The agent assists with evidence synthesis and calculation but cannot replace the nuanced clinical judgment of an experienced cardiologist, particularly in complex or atypical presentations.

23.3 Future Directions

1. Real-time ICU integration: Continuous hemodynamic monitoring with AI-driven trend analysis and early warning for cardiac decompensation.

2. Wearable data integration: Apple Watch ECG, continuous glucose monitoring, and remote blood pressure data feeding into longitudinal risk assessment.

3. Population health analytics: Cohort-level cardiovascular outcomes analysis across institutions using federated learning (NVIDIA FLARE).

4. Clinical trial matching: Automated screening of patients for cardiovascular clinical trial eligibility based on imaging, biomarker, and genomic criteria.

5. Cardiac digital twins: Patient-specific computational models combining imaging anatomy, electrophysiology, and hemodynamics for treatment simulation.

---

24. Conclusion

The Cardiology Intelligence Agent represents the natural next extension of the HCLS AI Factory platform into the highest-impact medical specialty. By leveraging proven multi-collection RAG patterns, shared infrastructure, and established NIM integration -- while adding cardiology-specific clinical workflows, validated risk calculators, GDMT optimization, and cross-modal genomic triggers -- the agent will deliver integrated cardiovascular intelligence that currently requires multi-million-dollar institutional investments.

The 12-collection Milvus architecture, 8 reference clinical workflows, 6 validated risk calculators, and GDMT optimization engine provide a comprehensive decision support platform that covers the full spectrum of cardiovascular practice: from preventive risk stratification through acute coronary syndromes to advanced heart failure management and cardio-oncology surveillance.

Deploying on a single NVIDIA DGX Spark ($3,999) ensures that this intelligence is accessible not just to elite academic centers but to community cardiologists, rural health systems, and resource-limited institutions worldwide. Combined with Apache 2.0 licensing and the HCLS AI Factory's proven open-source approach, the Cardiology Intelligence Agent will democratize cardiovascular AI in a way that no existing commercial product achieves.

---

25. References

1. Virani SS, et al. Heart Disease and Stroke Statistics -- 2025 Update. Circulation. 2025;151:e1-e375.
2. Heidenreich PA, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. Circulation. 2022;145:e895-e1032.
3. Writing Committee, et al. 2021 ACC/AHA/SCAI Guideline for Coronary Artery Revascularization. J Am Coll Cardiol. 2022;79:e21-e129.
4. Otto CM, et al. 2020 ACC/AHA Guideline for the Management of Patients with Valvular Heart Disease. Circulation. 2021;143:e72-e227.
5. January CT, et al. 2019 AHA/ACC/HRS Focused Update of the 2014 Guideline for Management of Atrial Fibrillation. Circulation. 2019;140:e125-e151.
6. Arnett DK, et al. 2019 ACC/AHA Guideline on the Primary Prevention of Cardiovascular Disease. Circulation. 2019;140:e596-e646.
7. Grundy SM, et al. 2018 AHA/ACC Cholesterol Clinical Practice Guideline. Circulation. 2019;139:e1082-e1143.
8. Lyon AR, et al. 2022 ESC Guidelines on Cardio-Oncology. Eur Heart J. 2022;43:4229-4361.
9. Vahanian A, et al. 2021 ESC/EACTS Guidelines for the Management of Valvular Heart Disease. Eur Heart J. 2022;43:561-632.
10. Hindricks G, et al. 2020 ESC Guidelines for the Diagnosis and Management of Atrial Fibrillation. Eur Heart J. 2021;42:373-498.
11. Ommen SR, et al. 2024 AHA/ACC Guideline for the Diagnosis and Management of Hypertrophic Cardiomyopathy. Circulation. 2024;149:e1239-e1311.
12. Cury RC, et al. CAD-RADS 2.0 -- 2022 Coronary Artery Disease Reporting and Data System. Radiology. 2022;305:209-221.
13. Lang RM, et al. Recommendations for Cardiac Chamber Quantification by Echocardiography in Adults: An Update from the ASE. J Am Soc Echocardiogr. 2015;28:1-39.
14. Kramer CM, et al. Standardized Cardiovascular Magnetic Resonance Imaging (CMR) Protocols: 2020 Update. J Cardiovasc Magn Reson. 2020;22:17.
15. Goff DC, et al. 2013 ACC/AHA Guideline on the Assessment of Cardiovascular Risk. Circulation. 2014;129:S49-S73.
16. Lip GY, et al. Refining Clinical Risk Stratification for Predicting Stroke and Thromboembolism in Atrial Fibrillation Using a Novel Risk Factor-Based Approach: The Euro Heart Survey. Chest. 2010;137:263-272.
17. Six AJ, et al. The HEART Score for the Assessment of Patients with Chest Pain in the Emergency Department. Neth Heart J. 2008;16:191-196.
18. Pocock SJ, et al. Predicting Survival in Heart Failure: A Risk Score Based on 39,372 Patients from 30 Studies (MAGGIC). Eur J Heart Fail. 2013;15:1082-1094.
19. Nashef SA, et al. EuroSCORE II. Eur J Cardiothorac Surg. 2012;41:734-744.
20. World Health Organization. Cardiovascular Diseases (CVDs) Fact Sheet. WHO, 2024.
21. FDA. Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. FDA Database, 2025.
22. McMurray JJ, et al. Angiotensin-Neprilysin Inhibition versus Enalapril in Heart Failure (PARADIGM-HF). N Engl J Med. 2014;371:993-1004.
23. McMurray JJ, et al. Dapagliflozin in Patients with Heart Failure and Reduced Ejection Fraction (DAPA-HF). N Engl J Med. 2019;381:1995-2008.
24. Packer M, et al. Cardiovascular and Renal Outcomes with Empagliflozin in Heart Failure (EMPEROR-Reduced). N Engl J Med. 2020;383:1413-1424.
25. Solomon SD, et al. Dapagliflozin in Heart Failure with Mildly Reduced or Preserved Ejection Fraction (DELIVER). N Engl J Med. 2022;387:1089-1098.
26. Maron DJ, et al. Initial Invasive or Conservative Strategy for Stable Coronary Disease (ISCHEMIA). N Engl J Med. 2020;382:1395-1407.

---

HCLS AI Factory -- Cardiology Intelligence Agent Research Paper and PRD Apache 2.0 License | March 2026 Author: Adam Jones